Do you put stable prompt rules before user-specific details to improve cache hits?

A practical question about keeping repeated AI API prompt prefixes stable enough for prompt caching to help.

Prompt caching sounds simple until the request shape keeps changing. I have seen teams put timestamps, user notes, and fresh search snippets before the stable instructions, then wonder why cached token counts stay low. A better pattern seems to be stable product rules, output format, examples, and tool schema first, then the changing user request at the end. That also makes prompt versions easier to compare. For API-heavy products, do you version the static prefix separately from the dynamic tail?